Page Q

Using Claude to write Codex PRDs, then review Codex quality. Plot twist: Claude started praising Codex. cc @OpenAI @OpenAIDevs @sama @gdb @kevinweil @nickaturley @fidjissimo @ryannystrom @derrickcchoi @varunrau #Codex #OpenAI #AIAgents #AgenticCoding #DevTools The original screenshot is Chinese, I use Codex to create an English version screenshot.

One more symptom: when asked what model it is, it may answer GPT-4 or GPT-5 nano instead of GPT-5.5 Pro / Extended. I also reproduced similar behavior in the desktop app, so this may not be web-only.

If anyone at OpenAI can check whether web Extended Pro is sometimes routing to fallback or dropping the thinking trace, that would be helpful.

Weird ChatGPT web issue with GPT-5.5 Pro / Extended Pro. Fresh chats + manual Pro + Extended briefly show "Pro thinking", then no persistent "Thought for ..."; output quality looks like a fallback/non-reasoning route. Thinking Heavy works normally. @OpenAI @ChatGPTapp @sama @nickaturley @barret_zoph Tried: new sessions, clearing chatgpt.com site data, relogin, disabling Fast answers. Also hit "Message delivery timed out. Please try again." My repro chats: chatgpt.com/c/69f63833-58f… chatgpt.com/c/69f63360-b4e… Similar reports: reddit.com/r/ChatGPTPro/c… reddit.com/r/ChatGPTPro/c… reddit.com/r/ChatGPTPro/c…

If anyone at OpenAI can check whether web Extended Pro is sometimes routing to fallback or dropping the thinking trace, that would be helpful.

Weird ChatGPT web issue with GPT-5.5 Pro / Extended Pro. Fresh chats + manual Pro + Extended briefly show "Pro thinking", then no persistent "Thought for ..."; output quality looks like a fallback/non-reasoning route. Thinking Heavy works normally. @OpenAI @ChatGPTapp @sama @nickaturley @barret_zoph Tried: new sessions, clearing chatgpt.com site data, relogin, disabling Fast answers. Also hit "Message delivery timed out. Please try again." My repro chats: chatgpt.com/c/69f63833-58f… chatgpt.com/c/69f63360-b4e… Similar reports: reddit.com/r/ChatGPTPro/c… reddit.com/r/ChatGPTPro/c… reddit.com/r/ChatGPTPro/c…

@nickaturley @alanxchen85 @OpenAI @ChatGPTapp I will send you a link

@alanxchen85 @OpenAI @ChatGPTapp That's odd. Any chance you could DM or share a convo link where this happened?

gpt 5.5 pro degrades to gpt 4 (!!!) on web, but is 5.5 pro on the desktop app? anyone else seeing this? @OpenAI @ChatGPTapp @nickaturley 1st in desktop app --> thinks 16s, tells me it's 5.5 pro (yeah that checks out, pro is slow) 2nd in browser --> instant response, is gpt-4

@alanxchen85 @OpenAI @ChatGPTapp @nickaturley Mac app doesn't work for me...

@qiaohui @OpenAI @ChatGPTapp @nickaturley Yeah so bizarre... but at least workaround w/ desktop app works but the app itself is a bit buggy sometimes 🤦‍♂️

@KevinQHLin @_akhaliq Great! Do you guys also grain an agent to play these games?

Thanks @_akhaliq sharing our work! Can frontier Multimodal Agents play games as well as humans? 🤩We are excited to introduce 🎮GameWorld: towards standardized and verfiable evaluation for multimodal game agents. 🕹️ 34 browser games 📌 170 tasks 🤖 18 multimodal agent baselines, covering 1. Computer-use (CUA) agents 👉 raw keyboard + mouse actions 2. Generalist multimodal agents 👉 semantic action parsinga GameWorld show that even sota agents still perform far below novice human players. 📹Watch our live runs: @gameworld-bench/live" target="_blank" rel="nofollow noopener">youtube.com/@gameworld-ben… 🌐project page: gameworld-project.github.io 💻github: github.com/gameworld-proj… work done with @OuyyyangMingyu @who_s_yuan Hwee Tou Ng, @MikeShou1

@steren @grok 请解释

WebMCP sits in a very interesting spot: It's not a spec for HTTP APIs, like MCP servers are. It's a way to help browsers with agentic capabilities use webapps more reliably (instead of asking them to understand the pixels on the screen or DOM) As a developer, you can reuse your existing JS logic, and transform it into WebMCP tools in just a few lines of JS. You can even transform any HTML form into a WebMCP tool by just adding new attributes.

@xf1280 Congrats! Gemini Robotics technical report provides a lot of details and very inspiring!

Excited that our work on Gemini Robotics and Gemini spatial understanding have just been featured on #GoogleIO stage! I believe that a frontier model possessing strong real-world understanding capabilities represents the ultimate path to embodied AGI, and we are making rapid advancements in this area.

@ericyi0124 Make sense, thanks! Another way to think about this question, when a robot hands something to human, how does the robot know the right time to release the object, the robot needs to make sure the human already hold the object.

@qiaohui We usually believe the later one is more challenging since robot have to perceive a half hidden object, handle human movement during hand over, and slip fingers in without contact.

Teaching mobile robots to safely fetch objects from a human—no real-world training needed 🚀 High-fidelity and reactive human simulation + safety-aware and imitation friendly demonstration planning + large scale imitation learning ⇒ zero-shot sim-to-real transfer. Chat with Zifan to know more on June 15, 3-5 PM at the main poster session! #CVPR2025

@tylerlum23 Great work, really inspiring!

🧑🤖 Introducing Human2Sim2Robot!  💪🦾 Learn robust dexterous manipulation policies from just one human RGB-D video. Our Real→Sim→Real framework crosses the human-robot embodiment gap using RL in simulation. #Robotics #DexterousManipulation #Sim2Real 🧵1/7

@chris_j_paxton Is there online video for your talk?

This is such a cool event, and its just too bad I'm not there in Singapore

@chris_j_paxton @vijayssundaram I always read your post and I think it's informative and less biased, keep posting your thoughts!

@vijayssundaram Robotics is not solved but there really is steady progress under all the hype. Personally I try to communicate this progress without the distortion

Social media doing what it does best: amplifying actually impossible fantasies over facts. Beliefs about AI (“robotics is solved”) are becoming as distorted as perceptions of others’ lifestyles. But the wildest part is watching researchers and VCs acting like vacuous celebrities.

@RemiCadene @LeRobotHF @huggingface Happy birthday!

Happy birthday @LeRobotHF Thanks to @huggingface community and team for your trust 🤗 Things are just getting started 🚀

Imitation learning for #humanoidrobots just got faster. 🥽 At #CES2025, the #NVIDIAIsaac GR00T blueprint was introduced for generating exponentially large synthetic datasets from just a small number of human demonstrations to train humanoids. Read Blog. nvda.ws/422wSPa